Fixes #20326

Type variables with forward references in upper bound are known to be
problematic. Existing mechanisms to work with them implicitly assumed
that they are used in the same module where they are defined, which is
not necessarily the case for "old-style" type variables that can be
imported.

Note that the simplification I made in `semanal_typeargs.py` would be
probably sufficient to fix this, but that would be papering over the
real issue, so I am making a bit more principled fix.

Fixes #20327

Fix is trivial, do not grab various internal/temporary symbols with star
imports. This may create an invalid cross-reference (and is generally
dangerous). Likely, this worked previously because we processed all
fresh modules in queue, not just the dependencies of current SCC.

Fixes #20093

This fixes the crash, but not the false positive (the false positive
existed prior to the regression that introduced the crash)

Fixes #20344

This PR fixes #20341

Fixes #20367

Fixes mypyc/librt#21

Fail with an explicit user-friendly error on PyPy.

Follow-up for #20384

Fixes #20353

This makes us respect e.g. `--output json` for cached files without
re-checking the files (which is the desired behavior for users, see
issue). This is also a first step towards resolving the "foo defined
here" conundrum for parallel checking.

The fix is straightforward. The only question was whether to continue
using `ErrorTuple`s or switch to a proper class. I decided to keep the
tuples for now to minimize the scope of change.

Note I am also adjusting generic "JSON" fixed-format helpers to natively
support tuples (unlike real JSON). We already use tuples in few other
places, so it makes sense to just make it "official" (this format is
still internal to mypy obviously).

Also change the header for 1.18 because of
#19910